微型电脑应用2025，Vol.41Issue(3)：10-14,5.
基于改进YOLOv5的复杂场景火灾检测方法研究
Research on Complex Scene Fire Detection Method Based on Improved YOLOv5
摘要
Abstract
Aimed at the lack of detection accuracy and speed of the current target detection algorithm in detecting flames and smoke in complex scenes,this paper proposes an algorithm model for fire detection in complex scenes based on improved YOLOv5 to improve the speed and accuracy of fire detection.Using YOLOv5 as the basic network,the Convolutional Block Attention Module(CBAM)attention mechanism is added to the backbone network,the characteristic information of flames and smoke without adding additional parameters is considered,and the redundant information is suppressed in the background.The(SepVit)module is used to replace the last layer of Spatial Pyramid Pooling-Fas(SPPF)in the backbone network.The model's flames and smoke feature extraction ability is improved.The Bidirectional Feature Pyramid Networks(BiFPN)structure is a-dopted to improve the enhanced Neck structure,and the model's flames and smoke feature fusion ability to improve the accura-cy of the flames and smoke detection.The positioning loss function is adjusted to Scaled Intersection over Union(SIOU)to im-prove model training speed and regression accuracy.Based on public data sets and Internet fire image data,10 080 flames and smoke data sample sets in multiple complex application scene are self-built,which solves the limitation of the lack of authorita-tive fire sample sets and improves the generalization ability of model training.The experimental results show that on the self-built flames and smoke dataset,the average accuracy of the improved algorithm is 7.1％higher than that of the original algo-rithm,and the detection speed reaches 96.8 frame/s.It can detect flames and smoke targets in complex fire scenes in real time.关键词
YOLOv5/火灾检测/CBAM/SIOU/SepVitKey words
YOLOv5/fire detection/CBAM/SIOU/SepVit分类
信息技术与安全科学引用本文复制引用
李井辉,汤伟业,刘一诺,王庆麒..基于改进YOLOv5的复杂场景火灾检测方法研究[J].微型电脑应用,2025,41(3):10-14,5.基金项目
国家自然科学基金项目(52274005) （52274005）
国家自然科学基金项目(52174001) （52174001）
